| Model | Verdict | Decode tok/s | Max ctx | Size | Chat | Coding | |
|---|---|---|---|---|---|---|---|
| Qwen/Qwen3.5-27B bartowski/Qwen_Qwen3.5-27B-GGUF · IQ2_XXS | Fits | 5.302.10–5.80 | 39K | 27.8B | p60 | p46 | |
| Qwen/Qwen3.8-27B unsloth/Qwen3.8-27B-GGUF · UD-IQ2_S | Fits | 5.902.40–6.50 | 39K | 27.8B | p59 | p33 | |
| Qwen/Qwen3.6-27B unsloth/Qwen3.6-27B-GGUF · UD-IQ2_XXS | Fits | 5.402.20–5.90 | 44K | 27.8B | p51 | p54 | |
| Qwen/Qwen3.5-9B unsloth/Qwen3.5-9B-GGUF · Q8_0 | Fits | 6.302.50–6.90 | 50K | 9.7B | p46 | – | |
| ornith-ai/Ornith-1.5-9B AtomicChat/Ornith-1.5-9B-GGUF · Q8_0 | Fits | 6.302.50–6.90 | 50K | 9.7B | p46 | p24 | |
| Qwen/Qwen3.5-4B unsloth/Qwen3.5-4B-GGUF · Q8_0 | Fits | 10.84.30–11.9 | 175K | 4.7B | p39 | – | |
| google/gemma-4-12B-it unsloth/gemma-4-12b-it-GGUF · Q6_K | Fits | 4.603.90–5.30 | 41K | 12B | p38 | – | |
| Qwen/Qwen3-4B-Instruct-2507 unsloth/Qwen3-4B-Instruct-2507-GGUF · Q8_0 | Fits | 7.906.70–9.10 | 43K | 4B | p38 | – | |
| ornith-ai/Ornith-1.0-9B unsloth/Ornith-1.0-9B-GGUF · Q8_0 | Fits | 6.302.50–6.90 | 50K | 9.4B | – | p20 | |
| prism-ml/Bonsai-27B-mlx-1bit MLX-1BIT | Fits | 8.503.40–9.30 | 87K | 26.4B | – | – | |
| ornith-ai/Ornith-1.5-9B-GGUF Q8_0 | Fits | 6.202.50–6.80 | 42K | 9.2B | – | – | |
| prism-ml/Ternary-Bonsai-27B-mlx-2bit MLX-2BIT | Fits | 5.902.30–6.40 | 38K | 26.4B | – | – | |
| Qwen/Qwen3-14B unsloth/Qwen3-14B-GGUF · IQ4_XS | Fits | 5.204.40–6.00 | 34K | 14.8B | – | – | |
| Qwen/Qwen2.5-Coder-14B-Instruct Qwen/Qwen2.5-Coder-14B-Instruct-GGUF · Q3_K_M | Fits | 5.204.40–6.00 | 32K | 14.8B | – | – | |
| Qwen/Qwen2.5-14B-Instruct Qwen/Qwen2.5-14B-Instruct-GGUF · Q3_K_M | Fits | 5.204.40–6.00 | 32K | 14.8B | – | – | |
| google/gemma-4-26B-A4B-it unsloth/gemma-4-26B-A4B-it-GGUF · UD-Q2_K_XL | Tight fit | 20.48.20–22.5 | 0 | 25.8B | p49 | – | |
| openai/gpt-oss-20b unsloth/gpt-oss-20b-GGUF · Q2_K | Tight fit | 20.48.20–22.4 | 0 | 21.1B | p39 | p15 | |
| Qwen/Qwen3-30B-A3B unsloth/Qwen3-30B-A3B-GGUF · UD-IQ1_M | Tight fit | 12.85.10–14.1 | 20K | 30.5B | – | – |
Quality = benchmark percentile against all indexed models (p50 = median). Speeds are single-user decode ranges.